Despite the high number of deep learning models presented in the last few years for automatically annotating medical images, clear baselines to compare model upon are still missing. Furthermore, though there are only two datasets publicly available for the task, there is neither a shared and commonly adopted procedure for preprocessing the raw data nor an unanimous way in which the intermediate tasks have been defined. The work here presented tries to fill this gap by clearly characterizing the datasets, defining the learning task and providing some baselines that can be especially helpful when trying to replicate the results in languages with less resources than those available in English.

Extended abstract, presented at the 1st International Serbian Conference on Applied Artificial Intelligence, 19-20 May 2022, Kragujevac, Serbia.